Associations between echocardiographic traits and AI-ECG predictions of heart failure

arXiv:2605.24576 · cs.AI · Submitted 2026-05-23 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Associations between echocardiographic traits and AI-ECG predictions of heart failure".

Jane: The gist The AI–ECG HF risk predictions align primarily with measures of systolic function, particularly global longitudinal strain, while also capturing diastolic-related abnormalities in patients with preserved LVEF.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So looking back at "Associations between echocardiographic traits and AI-ECG predictions of heart failure," the title itself sets up a direct comparison between the computer's prediction and established physical measurements of the heart.

Jane: The authors are essentially saying that their AI-ECG risk scores aren't just random guesses; they actually correspond to measurable cardiac problems we already understand from echocardiography, like strain and filling pressures <ref:2605.24576#pg1>

Lu: This suggests a path toward making AI diagnostic tools more trustworthy because we can use these physical traits to validate the model's predictions against known physiology <ref:2605.24576#pg1>

Meng: For us in production, this means we don't just trust the final risk number; we have a set of echo parameters that tell us *why* the AI thinks that way, which helps with clinical adoption and reducing false alarms <ref:2605.24576#pg1>

Lalam: The key finding is that the AIHF predictions primarily reflect left ventricular systolic function, specifically global longitudinal strain, while also capturing diastolic abnormalities in patients with preserved LVEF <ref:2605.24576#pg1>

Tom: So the implication for the broader field is that we can use these associations to improve model interpretability and refine the AI further by focusing on those specific physiological drivers <ref:2605.24576#pg1>

Jane: It’s about moving beyond just pattern recognition and getting a tool that speaks the language of cardiology, using strain as the primary guide for systolic risk <ref:2605.24576#pg1>

Conclusion: Tom: So, we’ve been looking at how AI uses ECGs to guess heart failure risk, and this paper is about checking if those guesses actually match what doctors see on an ultrasound or echo <ref:2605.24576#pg1>.

Jane: Yeah, the title says "Associations between echocardiographic traits and AI-ECG predictions of heart failure," which sounds complicated but basically they’re seeing if the computer’s risk score lines up with actual heart measurements.

Lu: The core idea is that AI can find things in the ECG that aren't always obvious, maybe not even captured by left ventricular ejection fraction, and this study tests if those hidden signals point toward real problems like remodeling or filling pressures.

Meng: So they’re taking a risk score from an AI model and correlating it with things we measure on an echo machine to see what drives that prediction. That sounds like a good way to check if the AI is learning the right stuff or just noise.

Lalam: Lalam sees that the main takeaway is that the AI HF predictions line up most strongly with systolic function markers, especially global longitudinal strain, but they also catch things related to filling pressures even in patients with preserved ejection fraction.

Tom: That makes sense, so it seems like the AI is picking up on those fundamental mechanical issues of the heart rather than just random electrical noise in the ECG signal.

Jane: Exactly—it’s suggesting that if we want to trust this AI, we need to look at these specific echoes parameters, like strain or diastolic indices <ref:2605.24576#pg1>.

Lu: The authors show that systolic function markers like GLS show the strongest correlation with the AI risk, and they also looked at subgroup differences by sex and LVEF.

Meng: I wonder how robust this is when you look at real-world deployment; they did external validation on a big dataset from Columbia University Irving Medical Center, which is good for checking generalizability.

Lalam: The external validation showed that the risk predictions held up across a much larger set of data points, confirming that these echo traits are consistent predictors of the AI's output.

Tom: So in conclusion, this study proves that we can use specific echoes measurements to interpret what an AI is predicting about heart failure risk, which should help make those models more clinically useful.

Medical Technology & E-Health, Akershus University Hospital · Faculty of Medicine, University of Oslo · Department of Cardiology, Akershus University Hospital · Institute of Clinical Medicine, Campus Ahus, University of Oslo

cs.AI

Submitted: 2026-05-23

Updated: 2026-05-23

DOI: 10.1093/ehjdh/ztag141

Code: https://github.com/Ahus-AIM/Associations-BetweenEchocardiographic-Traits-and-AI-ECG-Predictionsof-Heart-Failure

License: http://creativecommons.org/licenses/by-sa/4.0/

Importance score: 72/100

The gist: The gist The AI–ECG HF risk predictions align primarily with measures of systolic function, particularly global longitudinal strain, while also capturing diastolic-related abnormalities in patients

Key concepts

AI-ECG: This refers to using artificial intelligence on electrocardiogram (ECG) data to predict the risk of heart failure. The AI analyzes the heart's electrical signals from an ECG to flag potential HF cases, even when traditional measures like ejection fraction are normal.
Global Longitudinal Strain (GLS): GLS is a measurement derived from echocardiography that assesses the overall systolic function of the left ventricle across its entire length. It is considered one of the most accurate indicators of myocardial health and was found to be the strongest correlate with AI-ECG risk predictions.
Left Ventricular Ejection Fraction (LVEF): LVEF is a standard measure from echocardiography that shows what percentage of blood the left ventricle pumps out with each beat. The study examined how the AI predictions related to LVEF, finding that correlations were substantial for GLS and other measures specifically in patients with preserved LVEF.

Terminology

External Validation — 1× in this paper · explained in 2 episodes
Interpretability — 1× in this paper · explained in 15 episodes

Summary

The gist The AI–ECG HF risk predictions align primarily with measures of systolic function, particularly global longitudinal strain, while also capturing diastolic-related abnormalities in patients with preserved LVEF.

Background and Motivation

Artificial intelligence-enabled electrocardiography (AI–ECG) can detect heart failure (HF), including disease not captured by left ventricular ejection fraction (LVEF), but the cardiac phenotypes underlying model predictions remain unclear. To address this limitation, the study investigated whether AI–ECG-predicted HF risk aligns with established echocardiographic measures of myocardial dysfunction, remodelling, and filling pressures.

Methodology

The study retrospectively analyzed ECG and echocardiography data from 8,147 patients who underwent both examinations within three days at Akershus University Hospital between January 1, 2023, and June 1, 2025. A previously validated AI–ECG model for HF detection was applied to all ECGs. Spearman’s rank correlation ρ quantified associations between echocardiographic parameters and AI–ECG risk. Subgroup analyses were performed by sex and left ventricular ejection fraction (LVEF) External validation included 36,286 ECGechocardiography pairs from Columbia University Irving Medical Center.

Key Findings in the Internal Cohort

The results showed that Global longitudinal strain (GLS) showed the strongest correlation (ρ = 0.57), followed by mitral annular plane systolic excursion (MAPSE) (ρ = −0.49) and LVEF (ρ = −0.45) In patients with LVEF>50%, correlations remained substantial for GLS, MAPSE, and diastolicrelated parameters 50%, correlations remained substantial for GLS, MAPSE, and diastolicrelated parameters>. Volumetric left ventricular indices correlated less strongly in women, whereas diastolic indices showed stronger correlations in women than in men.

Subgroup and Sex Differences

Table 4 presented the Spearman rank correlation between AI-ECG HF risk prediction and echocardiographic parameters by LVEF subgroup and sex. After stratification by LVEF, GLS remained among the most strongly correlated indices. Diastolicrelated parameters exhibited higher correlations in patients with preserved LVEF. There were significant sex differences in LVEF and LVEDVi, with the model being more strongly correlated in men. Conversely, the model was significantly more correlated with E/e’, IVS, PWT, CI, and A-wave velocity in women than in men.

Clinical Interpretation and Conclusion

Physiological validation showed that AI–ECG HF risk predictions align primarily with measures of systolic function, particularly global longitudinal strain. This approach may improve clinical interpretability and identify opportunities for model refinement. The study concludes that the AI–ECG model’s HF predictions primarily reflect LV systolic function, with GLS showing the strongest association, followed by MAPSE and LVEF in men. The associations to diastolic and filling pressuresrelated parameters were most pronounced in patients with preserved LVEF, indicating that the model detects features consistent with HFpEF. Identifying the echocardiographic traits most strongly associated with predictions may help clinicians understand the physiological basis of model outputs. The present study should be interpreted with its limitations in mind.

External Validation and Generalizability

The same kernel-smoothing procedure was applied to the external EchoNext dataset collected at Columbia University Irving Medical Center, which consisted of 36,286 ECG-TTE pairs. Compared with the internal cohort, the external dataset exhibited a higher median model-predicted risk across the full range of echocardiographic parameters. The consistency supports the generalizability of using echocardiographic traits to interpret model behaviour at scale. The results may reflect differences in HF prevalence between groups, but this cannot be verified because diagnostic labels are not included in the CUIMC dataset.

Limitations of Explainability Methods

Current explainability methods for AI–ECG fall into two categories: inherent and post-hoc, each with limitations. Post-hoc approaches can nevertheless be useful in specific scenarios; when models operate on paper-ECG images, saliency maps may highlight regions unrelated to the task and thereby provide partial insight. Identifying the echocardiographic traits most strongly associated with predictions may help clinicians understand the physiological basis of model outputs. The present study should be interpreted with its limitations in mind.

Clinical Relevance

The model output increased sharply near clinically relevant thresholds, which aligns with how systolic and diastolic dysfunction is interpreted in routine practice. Patients with GLS of more than -18% consistently had low model-predicted risk, whereas patients with GLS of less than -16% had increasingly higher modelpredicted risk of HF. Patients with LVEF of above 55% also had low risk-predictions, whereas patients with LVEF below 40% had consistently high predicted risk. For PASP, values above 30 mmHg were correlated with increasing predicted risk.

Contributors

Elias Stenhede: Conceptualisation, Data curation, Formal analysis, Software, Visualisation, Writing - original draft Henrik Schirmer:

Improvements for AI systems

A system can be improved to provide physiological phenotyping of heart failure by incorporating established echocardiographic measures as predictors for AI–ECG model outputs, This approach enables the identification of HF phenotypes for which the model performs particularly well or poorly. This allows clinicians to understand what the model detects and highlights meaningful subgroup differences.
The AI system can be enhanced to better characterize heart failure with preserved ejection fraction (HFpEF) by focusing on diastolic function, as the associations to diastolic and filling pressures-related parameters were most pronounced in patients with preserved LVEF, indicating that the model detects features consistent with HFpEF. This enables the model to be more sensitive to subtle changes in ventricular relaxation and filling pressures.
The system can be made more robust by utilizing sex-stratified analysis to reveal emergent patterns, as the model appears to have learned sex-specific phenotypes, even though sex was neither a target nor a model input during training. This allows for the identification of potential biases or differences in model utility between sexes.
The system can be refined to offer more actionable clinical insights by leveraging specific thresholds, such as Patients with GLS of more than -18% consistently had low model-predicted risk, whereas patients with GLS of less than -16% had increasingly higher model predicted risk of HF. This provides concrete physiological benchmarks for risk stratification.
The system can be improved to enhance interpretability by focusing on the most strongly associated parameters, as the findings show GLS showed the strongest correlation (ρ = 0.57), followed by mitral annular plane systolic excursion (MAPSE) (ρ = −0.49) and LVEF (ρ = −0.45). This prioritizes feature importance based on established physiological validation.

Abstract

Artificial intelligence-enabled electrocardiography (AI-ECG) can detect heart failure (HF), including disease not captured by left ventricular ejection fraction (LVEF), but the cardiac phenotypes underlying model predictions remain unclear. We therefore investigated whether AI-ECG-predicted HF risk aligns with established echocardiographic measures of myocardial dysfunction, remodelling, and filling pressures. We retrospectively analysed ECG and echocardiography data from 8147 patients who underwent both examinations within three days at Akershus University Hospital between 1 January 2023 and 1 June 2025. A previously validated AI-ECG model for HF detection was applied to all ECGs. Spearman's rank correlation ρ quantified associations between echocardiographic parameters and AI-ECG risk. Subgroup analyses were performed by sex and left ventricular ejection fraction (LVEF). External validation included 36,286 ECG-echocardiography pairs from Columbia University Irving Medical Center. Global longitudinal strain (GLS) showed the strongest correlation (ρ =0.57), followed by mitral annular plane systolic excursion (MAPSE) (ρ =-0.49) and LVEF (ρ =-0.45). In patients with LVEF>50%, correlations remained substantial for GLS, MAPSE, and diastolic-related parameters. Volumetric left ventricular indices correlated less strongly in women, whereas diastolic indices showed stronger correlations in women than in men. Physiological validation showed that AI-ECG HF risk predictions align primarily with measures of systolic function, particularly global longitudinal strain, while also capturing diastolic-related abnormalities in patients with preserved LVEF. This approach may improve clinical interpretability and identify opportunities for model refinement.